Moonshot AI Unveils Kimi K3.1 Pro: A 2.8T MoE Multimodal Model with Open Weights Coming November 4
Moonshot AI has announced Kimi K3.1 Pro, a flagship multimodal model featuring 2.8T total parameters and a 1M token context window. API and web services are ava
Moonshot AI has officially announced Kimi K3.1 Pro, a flagship multimodal model featuring a total of 2.8T parameters (with 104B active MoE parameters) and a massive 1M token context window.
Image source: Moonshot AI (@Kimi_Moonshot)
Benchmark Performance and Capabilities
Kimi K3.1 Pro is engineered to excel in complex terminal coding and autonomous agent workflows.
- Terminal-Bench 4.0: Achieved 65.7% on realistic terminal coding tasks, outperforming GPT-6 Astra (58.2%) and coming close to Claude Opus 5.5 (66.4%).
- BenchCAD: Reached 92.7% accuracy with tool use on 3D reconstruction and CAD code generation benchmarks.
- Specialized Workflows: Scored 64.4 on OfficeQA Pro v2 and reduced hallucination rates by 65% in financial modeling. It also provides full-lifecycle support for Unreal Engine, Unity, Godot, Blender, and CAD development.
Service Availability and Open Weights Schedule
Alongside the announcement, Kimi K3.1 Pro and its 2–3x faster Highspeed variant are rolling out immediately.
- Ecosystem Access: Available now via the
kimi.aiweb interface, Kimi Work, Kimi Code, and the Kimi API. - Open Weights Release: Open weights for the model will be officially released on November 4, 2026.
- Caveats: Certain performance metrics—such as financial modeling improvements and business simulation gains—are based on Moonshot AI's internal benchmarks.
Sources
- Official Announcement: Moonshot AI X (@Kimi_Moonshot) Announcement